資料中心基礎設施
NVIDIA Spectrum-6 Pushes Single-Chip Ethernet Switching Capacity to 102.4 Tbps
NVIDIA has begun deploying Spectrum-6 switching systems designed for the Vera Rubin platform, delivering 102.4 Tbps of throughput on a single chip. The products also introduce 800Gb/s ports, liquid cooling, and co-packaged optics options, targeting scale-out networks with more than 100,000 GPUs.
NVIDIA announced that Spectrum-6 has entered the deployment phase in large-scale AI data centers. Its core switch chip delivers 102.4 Tbps of throughput, twice the capacity of the previous-generation system, and joins the ConnectX-9 SuperNIC, BlueField-4 DPU, Vera CPU, Rubin GPU, and NVLink 6 as part of the Vera Rubin platform. Rather than handling GPU interconnects within a rack, it carries scale-out traffic across servers and racks.
AI training and distributed inference frequently perform collective communications such as all-reduce and all-to-all. When large numbers of nodes transmit data simultaneously, conventional Ethernet designed for north-south enterprise traffic can suffer from congestion, tail latency, and path imbalance. The Spectrum-X hardware and software stack balances traffic across multiple paths, routes around failures, and performs precise retransmissions. NVIDIA claims that it can deliver up to 1.6× higher AI networking performance than standard Ethernet and maintain up to 95% network efficiency in deployments exceeding 100,000 GPUs; these figures currently remain vendor-reported test results.
Public hardware manuals show that both the SN6600 and the liquid-cooled SN6600-LD provide 128 800Gb/s OSFP ports and 256MB of packet buffering. Depending on the optical module configuration, the liquid-cooled SN6600-LD has typical power consumption of approximately 2.79–3.3kW. The larger SN6800-LD uses four Spectrum-6 ASICs to deliver 409.6 Tbps, with typical power consumption reaching 8.3kW. The products support pluggable or co-packaged optics, reflecting how optical I/O, cooling, and power delivery have become equally important system constraints as switching capacity increases.
Network engineers should next examine real-world collective communication completion times, tail latency during failure recovery, interoperability among different RDMA transport modes, and the maintenance costs of co-packaged optics. Spectrum-6 supports standard Ethernet and open network operating systems, but peak performance depends on the co-design of NVIDIA NICs, switch silicon, and software. Buyers must still assess whether the performance gains from this integration outweigh vendor lock-in and power-density concerns.